Real time systems including games have hard limits on how long something can take in the worst case, not the average. In games this manifests as: All the stuff has to be done in 16ms to render a frame on time, if we're late the user will perceive a glitch. The consequences can be worse in control systems in any number of real-world systems where a glitch can lead to stability problems.
So getting back to TFA, this is part of knowing what the "right" answer is for the context.
Yup, I can't remember which book but I'm pretty sure I remember reading some stuff from Michael Abrash about how some of the algorithms in quake were changed for worse average performance but better worst case performance. It's fairly intuitive when you think about the context... and that is the critical point - abstract problems are interesting, but it's important to evaluate them in their full context.
(Anywhere you're servicing some kind of interactive request might qualify, depending on this size of your datastructures.)
This is actually something I like to work through in interviews:
OK, you've got amortized-constant-time append to your array-backed colletion, but now what if we need it to be really constant time? How could we achieve this? (Same question & trick applies to hashtables.)
(The answer can be found, amongst other places, in the Go or redis source code.)
edit: or in pavpanchekha's comment, they beat me to it :-)
also, just realized you're the author of crafting interpreters. absolutely love that book!